computer-playbook/roles/web-app-stalwart

Stalwart Mail Server

Description

Runs Stalwart — a secure, all-in-one mail and collaboration server — as the platform's email provider. A single hardened service speaks SMTP, Submission, IMAP, POP3, JMAP, ManageSieve, CalDAV, CardDAV and WebDAV, with a built-in spam filter, DKIM/DMARC signing and a WebAdmin + REST management API.

This role replaces the end-of-life web-app-mailu role. Applications send mail through the same provider-agnostic abstraction (lookup('email') / sys-svc-mail) — no consumer changes are needed beyond the provider repoint.

Overview

Compared with Mailu's multi-container stack, Stalwart collapses the mail server into one binary. This role runs:

Container	Purpose
stalwart	SMTP/IMAP/JMAP/POP3/Sieve/DAV + built-in spam filter + WebAdmin/REST API
webmail (Roundcube)	Browser webmail (parity with Mailu)
clamav	Attachment antivirus, called as an SMTP DATA-stage milter
postgres (shared)	Account / mail / metadata store

Dynamic state (domains, accounts, passwords, DKIM, certificates) is administered at runtime via the JMAP management API; config.json only bootstraps the data store. Spam filtering is built into Stalwart; antivirus is provided by the clamav container (registered via x:MtaMilter, services.clamav.enabled — set it to false to run spam-only). Infected mail is rejected at DATA.

flowchart LR
    subgraph edge["Front proxy (TLS)"]
        MAIL_VHOST["mail.<domain> (WebAdmin)"]
        WEBMAIL_VHOST["webmail.<domain> (Roundcube)"]
    end

    subgraph stack["web-app-stalwart compose stack"]
        STALWART["stalwart<br/>SMTP/IMAP/JMAP/POP3/Sieve/DAV<br/>+ WebAdmin + REST API"]
        WEBMAIL["webmail (Roundcube)"]
        CLAMAV["clamav (DATA-stage milter)"]
        PG[("postgres<br/>(shared platform DB)")]
    end

    KC["Keycloak<br/>OIDC directory + SSO"]
    APPS["Platform apps<br/>(no-reply bot via msmtp)"]
    WORLD(("Internet<br/>ports 25/465/587/143/993/..."))

    MAIL_VHOST --> STALWART
    WEBMAIL_VHOST --> WEBMAIL
    WEBMAIL -- "IMAP 993 / SMTP 465<br/>XOAUTH2 (stalwart-webmail client)" --> STALWART
    STALWART -- "milter" --> CLAMAV
    STALWART --- PG
    STALWART -- "OIDC directory<br/>(token validation)" --> KC
    MAIL_VHOST -. "SPA login: PKCE<br/>(stalwart-webui client)" .-> KC
    WEBMAIL_VHOST -. "OAuth login" .-> KC
    APPS -- "SMTPS :465, AUTH<br/>(bot credential from the token store)" --> STALWART
    WORLD <-- "public mail ports<br/>(only when active MAIL_PROVIDER)" --> STALWART

Cosmos

The diagram places Stalwart Mail Server in the Infinito.Nexus cosmos: the components it deploys (capabilities), the central services it consumes (dependencies), and its outward reach (federation and bridged external networks).

flowchart LR
    subgraph deps [Dependencies]
        dep_svc_bkp_volume_2_local["svc-bkp-volume-2-local 💻"]
        dep_svc_db_postgres["svc-db-postgres 🐳🐝"]
        dep_svc_net_tor["svc-net-tor 🐳🐝"]
        dep_svc_net_tor_smtp["svc-net-tor-smtp 🐳🐝"]
        dep_web_app_dashboard["web-app-dashboard 🐳🐝"]
        dep_web_app_keycloak["web-app-keycloak 🐳🐝"]
        dep_web_app_matomo["web-app-matomo 🐳🐝"]
        dep_web_app_prometheus["web-app-prometheus 🐳🐝"]
        dep_web_svc_css["web-svc-css 💻"]
        dep_web_svc_logout["web-svc-logout 🐳🐝"]
    end
    subgraph role [web-app-stalwart 🐳🐝]
        svc_sso["sso"]
        svc_logout["logout"]
        svc_dashboard["dashboard"]
        svc_matomo["matomo"]
        svc_container_backup["container_backup"]
        svc_stalwart["stalwart"]
        svc_postgres["postgres"]
        svc_webmail["webmail"]
        svc_clamav["clamav"]
        svc_css["css"]
        svc_prometheus["prometheus"]
        svc_tor["tor"]
        svc_tor_smtp["tor-smtp"]
    end
    subgraph dependents [Dependents]
        dpt_sys_ctl_alm_email["sys-ctl-alm-email 💻"]
        dpt_web_app_akaunting["web-app-akaunting 🐳🐝"]
        dpt_web_app_baserow["web-app-baserow 🐳🐝"]
        dpt_web_app_bigbluebutton["web-app-bigbluebutton 🐳🐝"]
        dpt_web_app_bluesky["web-app-bluesky 🐳🐝"]
        dpt_web_app_bookwyrm["web-app-bookwyrm 🐳🐝"]
        dpt_web_app_bridgy_fed["web-app-bridgy-fed 🐳🐝"]
        dpt_web_app_checkmk["web-app-checkmk 🐳🐝"]
        dpt_web_app_confluence["web-app-confluence 🐳🐝"]
        dpt_web_app_decidim["web-app-decidim 🐳🐝"]
        dpt_web_app_discourse["web-app-discourse 🐳🐝"]
        dpt_web_app_erpnext["web-app-erpnext 🐳🐝"]
        dpt_more["..."]
    end
    dep_svc_bkp_volume_2_local -. "0..1" .-> svc_container_backup
    dep_svc_db_postgres -. "0..1" .-> svc_postgres
    dep_svc_net_tor -. "0..1" .-> svc_tor
    dep_svc_net_tor -. "0..1" .-> svc_tor_smtp
    dep_svc_net_tor_smtp -. "0..1" .-> svc_tor_smtp
    dep_web_app_dashboard -. "0..1" .-> svc_dashboard
    dep_web_app_keycloak -. "0..1" .-> svc_sso
    dep_web_app_matomo -. "0..1" .-> svc_matomo
    dep_web_app_prometheus -. "0..1" .-> svc_prometheus
    dep_web_svc_css -. "0..1" .-> svc_css
    dep_web_svc_logout -. "0..1" .-> svc_logout
    svc_stalwart -- "1:1" --> dpt_more
    svc_stalwart -- "1:1" --> dpt_sys_ctl_alm_email
    svc_stalwart -. "0..1" .-> dpt_web_app_akaunting
    svc_stalwart -. "0..1" .-> dpt_web_app_baserow
    svc_stalwart -. "0..1" .-> dpt_web_app_bigbluebutton
    svc_stalwart -. "0..1" .-> dpt_web_app_bluesky
    svc_stalwart -. "0..1" .-> dpt_web_app_bookwyrm
    svc_stalwart -- "0..0" --> dpt_web_app_bridgy_fed
    svc_stalwart -. "0..1" .-> dpt_web_app_checkmk
    svc_stalwart -- "0..0" --> dpt_web_app_confluence
    svc_stalwart -. "0..1" .-> dpt_web_app_decidim
    svc_stalwart -. "0..1" .-> dpt_web_app_discourse
    svc_stalwart -. "0..1" .-> dpt_web_app_erpnext
    linkStyle 18,20 stroke:red;

Solid 1:1 edges are fixed relationships; dashed 0..1 edges are conditional (enabled only in matching deployments); red 0..0 edges are turned off in this role. Node markers show the role's deploy modes (💻 host, 🐳 compose, 🐝 swarm); ❌ marks a service that is explicitly turned off, and ⚙️ an Ansible role dependency declared in meta/main.yml.

SSO (Keycloak / OpenID Connect)

When web-app-keycloak is present the role joins the shared Keycloak client and delegates interactive authentication to Keycloak:

• WebAdmin + IMAP/SMTP/JMAP authenticate against an OIDC directory (tasks/06_oidc.yml): Stalwart validates Keycloak-issued tokens, matching the preferred_username claim to the provisioned account.
• Roundcube logs users in over OAuth2 and talks to Stalwart with XOAUTH2 (templates/roundcube-oauth.inc.php.j2).
• The OIDC directory skips the audience check and gives Keycloak tokens without Stalwart scopes access to mail, contacts, calendars and files (STALWART_OIDC_TOKEN_VALIDATION). Tokens of the stalwart-webui client get full access, because the WebUI Application names it as its oauthClientId.

Design constraint (validated against the live JMAP schema): Stalwart's authentication directory is Internal XOR one external directory — there is no chaining or fallback. Enabling SSO therefore disables the account password for every account, including the machine no-reply account. App passwords are checked before the directory and keep working, so with SSO enabled the role:

1. Creates one app password per bot account, acting for that account as the recovery admin (tasks/04_manage_user/app_password.yml), and
2. Stores it in the token store, where plugins/lookup/email.py reads the SMTP password of the no-reply client.

A bot gets a new app password when its entry named STALWART_BOT_APP_PASSWORD_DESCRIPTION is deleted in Stalwart or its token is removed from the token store. The app password it replaces is renamed to STALWART_BOT_APP_PASSWORD_RETIRED_DESCRIPTION and stays valid until the service-loader post-load flush drops it (tasks/04_manage_user/app_password_drop.yml), or until right after the new one is stored when the role runs outside the service loader. A deploy that stops before the flush leaves it to the next deploy. Relaying stays limited to authenticated sessions in both modes (STALWART_RELAY_RULE).

With SSO disabled the role uses the Internal directory and the bot's derived account password.

WebUI sign-in

The WebUI Application always names stalwart-webui (STALWART_WEBUI_OAUTH_CLIENT_ID) as its oauthClientId, with and without SSO. Where the WebUI sends a login depends on the account:

Account	Authorization server	Client stalwart-webui
Account in a domain bound to the Keycloak OIDC directory	Keycloak	Keycloak client from templates/keycloak/clients.json.j2
Recovery admin (STALWART_ADMIN_USER), and every account while SSO is off	Stalwart	OAuth client the role registers in Stalwart (tasks/02_webadmin.yml)

The role registers that Stalwart client over JMAP (templates/jmap/oauth_client/) as a public first-party client, with one redirect URI <origin>/<prefix>/oauth/callback and one post-logout URI <origin>/<prefix>/login per URL prefix of the WebUI Application. <origin> is the role's canonical WebUI URL (STALWART_WEBUI_BASE_URL), so a Tor deployment registers its http:// onion origin. Every deploy compares the registered client with that target and updates it when they differ. A deployment that serves the WebUI on a further origin MUST add that origin's URIs to templates/jmap/oauth_client/webui.json.j2, because a URI added in Stalwart alone is removed by the next deploy. Stalwart's login page asks for consent before the password form for every registered client, so a native WebUI login shows an Allow step.

Calendar & Contacts (CalDAV / CardDAV / WebDAV)

Stalwart serves DAV natively on the mail HTTP listener — no extra container (Mailu used a separate Radicale service). Clients auto-discover via https://mail.<domain>/.well-known/{caldav,carddav}, which redirect to https://mail.<domain>/dav/cal and /dav/card. Authenticate with the mailbox account (or the Keycloak SSO token when SSO is enabled).

Design Decisions

• Client ports are implicit-TLS only (465/993/995). Stalwart's default configuration binds no STARTTLS listeners, so the role neither publishes 587/143/110 nor advertises them via SRV records.
• The WebUI (WebAdmin) bundle is staged by the controller. Stalwart fetches it from GitHub on first boot, which fails behind restricted egress; the controller downloads it into a host-level cache (matrix rounds purge the instance dir, and per-round downloads hit GitHub rate limits) and bind-mounts it read-only. The update: block next to webui_version in meta/services.yml names the upstream tags, so the version-source updater bumps the bundle on its own, within the v2. line that server 1.x serves.
• The embedded PostgreSQL pins the plain postgres image — the platform's postgis default defeats baudolo's substring database detection and would degrade backups to torn live-file snapshots. backup.project_hard_restart brings the whole project back through one coherent restart after baudolo's per-volume stop/start cycles.
• ClamAV runs fail-open with short milter timeouts so mail keeps flowing while clamd warms its signature database instead of blocking the DATA stage.
• DMARC aggregate reports go to the mailbox of services.stalwart.dmarc.rua_user (the administrator user unless the inventory names another user key). The _dmarc record publishes rua=mailto:<username>@<mail domain>, and a deploy whose report user carries no mailbox account fails before any record is written.
• config.json is bootstrap-only (data store selection); all other configuration lives in the database and is provisioned over JMAP (templates/jmap/*.json.j2).
• Only authenticated sessions relay, with or without SSO. Under SSO a bot authenticates with an app password.
• Address bans expire after services.stalwart.ban_period_seconds (one hour by default). Without a period Stalwart keeps a ban forever, and the reverse proxy, the published ports, Tor and the swarm ingress hand many clients one address.
• Bans and failed logins also go to the container output (STALWART_BAN_TRACER, tasks/tracing.yml), next to the daily log file in the data volume.

Features

• All-in-one mail server (SMTP/IMAP/JMAP/POP3/ManageSieve)
• CalDAV / CardDAV / WebDAV collaboration
• Built-in spam filtering
• ClamAV antivirus as an SMTP DATA-stage milter (services.clamav.enabled: false for spam-only)
• DKIM signing with automatic key management; SPF / DMARC published in DNS
• OpenID Connect SSO via Keycloak
• Roundcube webmail
• PostgreSQL backing store

Quick Setup

Development

Clone, set up the workstation, and deploy Stalwart Mail Server onto the local stack:

git clone https://github.com/infinito-nexus/core.git
cd core
make onboard
make compose-deploy mode=reinstall apps=web-app-stalwart full_cycle=false

Production

Run the published image to provision the inventory and deploy Stalwart Mail Server to a managed server (the mounted volume persists the inventory):

APP=web-app-stalwart
HOST=<your-server>
DOMAIN=<your-domain>
TLS_MODE=self_signed
SSH_PUBLIC_KEY="<your-ssh-public-key>"

docker run --rm -it \
  -v "$PWD/inventories:/etc/infinito.nexus/inventories" \
  -e APP="$APP" -e HOST="$HOST" -e DOMAIN="$DOMAIN" -e TLS_MODE="$TLS_MODE" -e SSH_PUBLIC_KEY="$SSH_PUBLIC_KEY" \
  ghcr.io/infinito-nexus/core/debian bash -c '
    INVENTORY=/etc/infinito.nexus/inventories/production
    infinito administration inventory provision "$INVENTORY" \
      --inventory-file "$INVENTORY/devices.yml" \
      --host "$HOST" \
      --include "$APP" \
      --vars "{\"TLS_MODE\": \"$TLS_MODE\", \"DOMAIN_PRIMARY\": \"$DOMAIN\", \"users\": {\"administrator\": {\"authorized_keys\": [\"$SSH_PUBLIC_KEY\"]}}}" &&
    infinito administration deploy dedicated "$INVENTORY/devices.yml" \
      --password-file "$INVENTORY/.password" \
      --diff -vv'

Migration from Mailu

The provider cutover is an inventory change; mailbox data moves with it when the migration switch is on.

Compose only. The import reads the legacy maildir volume on the stack host and submits over the host-published IMAPS port, neither of which holds under swarm. Migrate in compose mode, then move the stack to swarm.

1. Keep web-app-mailu in the host's groups and add web-app-stalwart — both MUST be present so Mailu re-renders as a legacy instance (legacy-mail.<domain>, no public mail ports) and releases mail.<domain> to Stalwart.
2. Remove MAIL_PROVIDER from the inventory (Stalwart is the default).
3. Set applications["web-app-stalwart"].services.stalwart.migration.import_mailu: true and run the full deploy. The role then reads Mailu's Dovecot Maildir volume and imports every inventory account's messages via IMAP APPEND, preserving folders, flags and internal dates. Re-runs are idempotent (Message-ID dedup), so the switch may stay on across deploys; turn it off once Mailu is retired. The import authenticates per account with the inventory password, so it MUST run while services.sso.enabled is false. An OIDC directory replaces password login outright (see tasks/06_oidc.yml), and every IMAP login then fails with AUTHENTICATIONFAILED. Migrate first, enable SSO afterwards.
4. Accounts and aliases are provisioned from the inventory by this role; mailboxes created only inside Mailu's admin UI MUST be added to the inventory first. Sieve filters and CalDAV/CardDAV data are out of scope.
5. Non-Cloudflare DNS: publish the new DKIM TXT record reported by the deploy; MX and A records keep their hostname.

The cutover is covered end to end by files/test/test.sh, which runs as this role's CLI test whenever web-app-mailu is co-deployed — matrix variant 1, the round where SSO is off, because the import authenticates with per-account passwords. The whole scenario runs inside the deployed host and needs nothing from the environment around it: it stores mail in both directions in the legacy Mailu, confirms it landed in the maildir volume, runs files/python/migrate_from_mailu.py — the same script and argv a real cutover uses — and then asserts over IMAP that the stored mail survived and that live mail still flows. No redeploy is involved; the cutover is a docker-level data move.

Upgrading from 0.16

Stalwart 1.x refuses to open a 0.16 store and has no in-place upgrade. Every deploy checks the store in the role's database before it starts Stalwart, whatever state the running server is in, and stops when the store predates 1.0 (tasks/store_guard.yml). A store from before 1.0 keeps a bigint column v in table u, a 1.x store a bytea one, and a database without that column holds no store yet.

A deploy that has brought Stalwart 1.x up on the store and finished its provisioning writes the marker store.layout, holding the store layout version 1, into the role's instance directory on the stack host (STALWART_STORE_MARKER_FILE, written last in tasks/manager_ops.yml). A failed deploy writes no marker. A deploy that finds the marker passes without reading the store. A deploy without the marker reads the store:

Database	No store yet	Store to read
Shared PostgreSQL	the postgres superuser finds no database stalwart in the catalog	read as the role's database user
Dedicated PostgreSQL	the volume stalwart_database does not exist	read inside the dedicated database container; when that container is not running, the deploy stops and names the database service to start

Migrate the store with stalwart-migrate into a new database and swap it in under the role's database name. That name is fixed to stalwart (plugins/lookup/database.py derives it from the entity name), so the migration targets a second database that later takes over the name.

1. Make sure the server runs Stalwart v0.16.27; stalwart-migrate reads no other release. Upgrade an older 0.16 server with the previous revision of this role first.

2. Back up the database stalwart (for example pg_dump -Fc -d stalwart -f stalwart-v016.dump) and the role's volumes.

3. As a PostgreSQL superuser, create the empty target database, owned by the role's database user:

   CREATE DATABASE stalwart_v1 OWNER stalwart;
   

4. Copy the rendered config.json (STALWART_CONFIG_FILE, etc/config.json in the role's volumes directory) into a new directory and set its "database" to "stalwart_v1". This file is the --target.

5. Answer the conversion questions and copy the data while 0.16 keeps serving. Run the tool on the network the stalwart container reaches its database on (docker inspect stalwart), so the hosts in both files resolve:

   docker run --rm -it --network <stalwart-network> \
     -v <volumes-dir>/etc:/etc/stalwart:ro \
     -v <target-dir>:/etc/stalwart-v1 \
     -v <state-dir>:/var/lib/stalwart-migrate \
     stalwartlabs/stalwart-migrate:1.0.0 prepare \
       --source /etc/stalwart/config.json \
       --state /var/lib/stalwart-migrate
   

   Then run the same container with run --source /etc/stalwart/config.json --target /etc/stalwart-v1/config.json --state /var/lib/stalwart-migrate. run copies only what changed since the previous run and may be repeated.

6. Stop the Stalwart server (docker stop stalwart in compose, docker service scale <stack>_stalwart=0 in swarm) and run the same container with run --final and the same options. It verifies the copy and writes final-report.txt into the state directory.

7. Swap the databases, as a PostgreSQL superuser connected to the postgres database:

   ALTER DATABASE stalwart RENAME TO stalwart_v016;
   ALTER DATABASE stalwart_v1 RENAME TO stalwart;
   

8. Deploy this role. The store guard finds the migrated 1.x store under the role's database name and lets Stalwart 1.x start on it.

9. Keep stalwart_v016, the dump and the state directory until 1.x is verified. stalwart-migrate rollback reads the state directory to copy the 1.x store back into a new 0.16 store.

Restore and redeploy

The marker lives outside the stack and its volumes, so it survives docker stack rm, docker compose down, a crashed database and the stop and start cycle of backup.project_hard_restart. A host that carries the marker therefore redeploys while its dedicated database is down, as in the restore of svc-bkp-nfs-2-local (docker stack rm, restore, redeploy). A dedicated store on a host without the marker, such as a 0.16 host or a host whose instance directory was lost, needs its database running before the deploy: start the database service of the stack the host last deployed, then deploy. Before a host that carries the marker goes back to a 0.16 store, through stalwart-migrate rollback or a backup taken under 0.16, delete the marker, so the next deploy reads the store.

Verifying a production server

Deploying against a publicly reachable mail server additionally asserts that server's live hard facts — the records and listeners either match the deployed configuration or they do not. There is no mail-flow scenario here; that is what the suites above are for.

Set INFINITO_MAIL_PRODUCTION_HOST to the server's public FQDN (in CI, a GitHub repository variable of the same name; empty disables the whole check). Optionally set INFINITO_MAIL_PRODUCTION_RESOLVER to the address of a public resolver so the records are read from the outside view instead of /etc/resolv.conf, which matters on a host whose own resolver answers from a split-horizon zone.

files/python/production_facts.py then checks, read-only:

Fact	Assertion
A / AAAA	the mail host resolves
MX	the zone's only MX is the mail host (the role publishes it solo)
SPF	exactly one v=spf1 record, authorising mx or a:<mailhost>
DMARC	_dmarc publishes a v=DMARC1 record with a policy whose rua names the deployed aggregate report mailbox
DKIM	the selector this deploy read back publishes a non-empty v=DKIM1 key
SRV	_submissions/465, _imaps/993, _pop3s/995 point at the mail host
PTR	rDNS does not contradict the mail host (absent → warning; rDNS is provisioned for Hetzner only)
SMTP	port 25 greets with 220 carrying the FQDN, proving the rejectNonFqdn identity

The same variable points sys-ctl-mtn-cert-deploy's remote check at :465 and :993, so the live TLS chain and its remaining validity are verified alongside. A misconfigured production server therefore fails the deploy instead of surfacing as an outage.

Further Reading

• Stalwart documentation
• sys-svc-mail — how applications send mail
• plugins/lookup/email.py — the email abstraction

Note: Stalwart's JMAP object schema is version-sensitive. Pin services.stalwart.version; the provisioning payloads in tasks/ were validated against that release's GET /api/schema.

Credits

Implemented by Alejandro Roman Ibanez. Part of the Infinito.Nexus Project and maintained by Kevin Veen-Birkenbach. Licensed under the Infinito.Nexus Community License (Non-Commercial).